MCP Node

An MCP node connects an AI agent to an external tool server — a service that exposes a set of callable functions over the Model Context Protocol. Wire one into an Agent and the agent can call that server's tools mid-answer: look up library docs, search a wiki, query a database, hit a GitHub repo, and so on. The agent decides when to call them based on what you ask it to do.

Quick start

1. Register the server once in Settings → MCP → Add Server. Give it a name, pick a transport, and paste the endpoint URL (or, for a local command-line server, the command + args). See Adding a server below.
2. Drag an MCP node from the side panel (the plug icon, between Store and Sheet).
3. Pick your registered server from the node's Server dropdown.
4. Draw a wire from the MCP node into an Agent node.
5. In the agent's prompt, tell it to use the server for the relevant task (see below).
6. Run the agent. It connects to the server, sees the available tools, and calls the right one.

What to type in the agent

You don't need to name specific functions or learn an API. Just tell the agent, in plain English, to use the connected server for a task:

Use Context7 to find the current setup steps for Next.js server actions,
then summarise them.

Use DeepWiki to read the structure of the "facebook/react" repo and list
its main topics.

Use the GitHub tools to find the most recently updated open issue in our repo.

The pattern "use [server] to [do some task]" is almost always enough. The agent reads each tool's description (provided by the server itself) and picks the right one, fills in the arguments, and calls it. You generally don't need to mention tool names by hand — though you can if you want to force a specific one.

Tip: if the agent answers from its own knowledge instead of calling the server, make the instruction more explicit — "Use the Context7 tools — don't answer from memory" — or pick a more tool-eager model (GPT-4o Mini is a good default).

Filtering which tools the agent sees

Some servers expose dozens of functions. You usually don't want all of them in the agent's context — it's slower and the agent can pick wrong. Open the MCP node's config panel (the gear icon on the node header) to control this:

• All tools (default) — every tool the server advertises is exposed to the agent.
• Untick All tools, then tick only the functions you want. The agent sees just those.
• Tools are grouped by name prefix where the server uses one (e.g. repo_*, issue_*). Use the group checkbox to toggle a whole group at once, or expand a group for individual control.
• Refresh re-fetches the tool list from the server (use it if the server added or changed tools).
• The 📋 copy icon next to any tool copies its name — handy if you want to reference a specific function explicitly in the agent prompt.

Keeping each MCP node focused on a few relevant tools is the recommended pattern — it's cleaner than one node exposing everything.

Multiple servers on one agent

You can wire several MCP nodes into the same agent — one for docs, one for web search, one for your database. Each server's own usage guidance is added to the agent's prompt, so the agent can tell them apart ("Context7 for library docs, DeepWiki for repo wikis"). Just describe the task and the agent routes to the right server.

If two servers happen to expose a tool with the same name, Circuitry automatically disambiguates them so there's no clash.

Adding a server

In Settings → MCP → Add Server:

• Name — a label you'll recognise (also shown on the node).
• Transport:
  • URL — Streamable HTTP (recommended) — for hosted servers like https://mcp.example.com/mcp.
  • URL — SSE — the legacy transport, for older deployments.
  • Local command — stdio — runs a command-line MCP server (e.g. npx -y @modelcontextprotocol/server-filesystem /tmp) as a child process of your EServer. Needs an EServer running.
• Endpoint URL — for the URL transports.
• Headers (optional) — for servers that need an API key, e.g. Authorization: Bearer YOUR_TOKEN, one per line.

You can copy/paste your whole server list between machines using the copy/paste icons at the top of the MCP settings section — handy for moving setups between browsers or devices. Pasting skips any servers you already have.

Seeing what happened

When the agent runs, the Run Log (top toolbar) shows the call flow:

agent  AI Agent          ← the agent starts
  mcp  Context7          ← connects, lists tools
  mcp  Context7 · query-docs   ← a real tool call
agent  AI Agent  4.4s    ← the agent's final answer

Tool calls are indented under the agent that made them. The edge between the MCP node and the agent animates while a call is in flight. Click any row to see the raw tool output.

Troubleshooting

• "pick a server" on the node — open the gear panel and choose a registered server from the dropdown.
• No tools listed in the gear panel — the node connects when you open the panel; if it stays empty, press Refresh, or check the server is enabled and reachable in Settings → MCP.
• Agent ignores the server — make the prompt explicit ("use [server] to …, don't answer from memory"), or switch to a tool-capable model.
• Connection times out — the node surfaces an error on its subtitle and in the Run Log within ~15s. Check the endpoint URL and any required auth headers in Settings.
• stdio server won't start — make sure an EServer is running and the command is installed on that machine; the error message in the Run Log says what failed.